Tag
I grew a resilience audit from a single agent into a scatter-gather graph with typed edges, an adversarial verification gate and injected failures. Recall barely moved; trust did.
What happens when an AI model returns typed decisions and probabilities instead of generating an answer? A look at Jev, calibration, and a possible fast path inside the agent runtime.
A 5 minute experiment: wiring Alibaba's page-agent into my career site and pointing it at DeepSeek, so the whole UI takes orders in plain English. How it works under the hood, what its reasoning traces actually look like, and why I kept it strictly local.
An evening climbing the loop ladder in Claude Code, from a one-shot fix to an unattended cloud routine. The lessons had nothing to do with the code and everything to do with how a loop stops: unenforceable budgets, self-writing reward hacks, and why a trapped loop cannot free itself.